본문으로 건너뛰기

Core Concepts

FluxLoop is built on a few key concepts that enable robust agent testing at scale.

Agent​

An agent is any function that processes inputs and produces outputs. In FluxLoop, you mark agent entry points with the @fluxloop.agent() decorator:

import fluxloop

@fluxloop.agent()
def my_agent(prompt: str) -> str:
    return process(prompt)

The decorator automatically captures execution data, timing, and results.

Trace​

A trace is the complete record of a single agent execution. It captures:

  • Inputs: What the user said
  • Outputs: What the agent replied
  • Steps: Internal tool calls, LLM queries, and state changes
  • Metrics: Latency, token usage, and estimated cost

Persona​

A persona is a synthetic user archetype. Instead of just testing "How do I start?", you test it from the perspective of a "frustrated novice" or a "technical power user". This helps uncover edge cases in how agents handle different tones and levels of expertise.

Scenario​

A scenario is a curated test case that combines personas, inputs, and expected outcomes. Scenarios can be:

  • Local: Defined in your fluxloop.yaml or YAML files.
  • Cloud-managed: Centrally managed on the Web Platform for team-wide consistency.

Test Run​

A test run is a batch execution of one or more scenarios. When you run fluxloop test, you create a local test run. When you upload it, it becomes a permanent record on the Web Platform.

Input Generation​

FluxLoop uses LLMs to generate realistic input variations. From a single base input like "Where is my order?", FluxLoop can generate 100 variations like:

  • "Can you tell me where my package is?" (Polite)
  • "Still waiting for my delivery, any updates?" (Impatient)
  • "Order #12345 status please." (Concise)

Web Platform​

The Web Platform (app.fluxloop.ai) is the central hub for:

  • Visualization: Interactive dashboards for exploring traces.
  • Collaboration: Sharing results and scenarios with your team.
  • Evaluation: Automating pass/fail checks based on criteria.
  • History: Tracking agent performance over time.

Workflow​

The typical FluxLoop workflow is "Agent-First":

  1. Build: Write your agent and add @fluxloop.agent().
  2. Test: Run /fluxloop test in Claude Code or fluxloop test in CLI.
  3. Analyze: View results on results.fluxloop.ai.
  4. Iterate: Improve your agent based on insights and re-test.

Framework Agnostic​

FluxLoop works with any agent architecture:

  • LangChain / LangGraph: Automatic integration with native tracing.
  • Custom Code: Direct use of decorators and SDK.
  • API-based Agents: Test via HTTP or other interfaces.

Next Steps​